Workflow Manager

Workflow Manager Skill

Multi-agent workflow orchestration for coordinating Claude, Cursor, and Gemini agents.

Overview

This skill enables Claude Code to act as the lead orchestrator in a multi-agent development workflow. Claude coordinates with Cursor (code quality, security) and Gemini (architecture, scalability) to implement features with robust validation and verification.

Usage

/workflow-manager

Prerequisites

• SurrealDB configured for workflow state persistence.

Capabilities

Core Orchestration

• Initialize and manage 5-phase workflows
• Track workflow state with persistence and resumability
• Coordinate parallel agent execution
• Handle context versioning and drift detection

Approval Management

• Evaluate agent feedback using configurable policies
• Support multiple approval strategies: ALL_MUST_APPROVE, NO_BLOCKERS, WEIGHTED_SCORE, MAJORITY
• Track iteration counts for plan-validate cycles

Conflict Resolution

• Detect conflicts between agent feedback
• Resolve using weighted expertise areas
• Support escalation to human decision when needed

Workflow Phases

File Structure

project/
├── AGENTS.md           # Workflow rules (source of truth)
├── PRODUCT.md          # Feature specification
├── CLAUDE.md           # Claude-specific context
├── GEMINI.md           # Gemini context
├── .cursor/rules       # Cursor rules
├── scripts/
│   ├── call-cursor.sh  # Cursor CLI wrapper
│   └── call-gemini.sh  # Gemini CLI wrapper
└── (SurrealDB)          # workflow_state, phase_outputs, logs

Agent Expertise Areas

Cursor (Code Quality Focus)

• Security: 0.8 weight - SQL injection, XSS, OWASP vulnerabilities
• Code Quality: 0.7 weight - Style, best practices, maintainability
• Testing: 0.7 weight - Test coverage, test quality

Gemini (Architecture Focus)

• Architecture: 0.7 weight - Design patterns, modularity
• Scalability: 0.8 weight - Performance at scale, bottlenecks
• Patterns: 0.6 weight - Design patterns, anti-patterns

Approval Policies

Phase 2 (Validation) - NO_BLOCKERS

• Approve if no high-severity blocking issues
• Minimum combined score: 6.0
• Single agent can approve if other unavailable

Phase 4 (Verification) - ALL_MUST_APPROVE

• Both Cursor and Gemini must approve
• Minimum score: 7.0
• No blocking issues allowed
• Uses CONSERVATIVE conflict resolution

Context Versioning

The workflow tracks checksums of context files:

• AGENTS.md - Workflow rules
• PRODUCT.md - Feature spec
• .cursor/rules - Cursor rules
• GEMINI.md - Gemini context
• CLAUDE.md - Claude context

Drift detection warns when files change mid-workflow.

Commands

Usage Example

User: Implement the feature in PRODUCT.md

Claude: I'll start the multi-agent workflow.
[Reads AGENTS.md, PRODUCT.md]
[Creates plan.json in Phase 1]
[Runs /validate-plan - Cursor and Gemini review in parallel]
[Implements with TDD in Phase 3]
[Runs /verify-code - Both agents must approve]
[Completes Phase 5 with documentation]

Integration

With Python Orchestrator

from orchestrator import Orchestrator

orch = Orchestrator("/path/to/project")
orch.run()  # Runs all 5 phases

With Bash Scripts

# Initialize project
bash scripts/init-multi-agent.sh /path/to/project

# Call agents directly
bash scripts/call-cursor.sh prompt.md output.json
bash scripts/call-gemini.sh prompt.md output.json

State Management

The workflow state is persisted in SurrealDB (workflow_state table):

{
  "project_name": "my-project",
  "current_phase": 3,
  "iteration_count": 1,
  "phases": {
    "planning": {"status": "completed"},
    "validation": {"status": "completed"},
    "implementation": {"status": "in_progress"}
  },
  "context": {
    "files": {"agents": {"checksum": "abc123..."}}
  }
}

Error Handling

• Phases can retry up to 3 times (configurable)
• Blockers are recorded and reported
• Failed phases can be resumed after fixing issues
• Context drift is detected and logged

Best Practices

1. Read AGENTS.md first - Contains complete workflow rules
2. Follow TDD in Phase 3 - Write tests before implementation
3. Address all blocking issues - Required for approval
4. Don't skip phases - Each phase builds on previous
5. Check drift regularly - Use /phase-status to monitor

Outputs

• Updated workflow records in SurrealDB (workflow_state, phase_outputs, logs).

Related Skills

• /orchestrate - Run the full workflow
• /phase-status - Detailed phase status